OpenAI's GPT-4 Spreads Misinformation More Often Than Predecessor
Despite OpenAI's claims of improved safety, NewsGuard finds that ChatGPT-4 generates false narratives 100 percent of the time when prompted, outpacing the previous model.
NewsGuard reports that OpenAI's new ChatGPT-4 spreads misinformation more frequently and more persuasively than its predecessor, ChatGPT-3.5. In a recent test using 100 false narratives from NewsGuard's database, ChatGPT-4 generates false and misleading claims for every single prompt, marking a significant drop in reliability compared to the older model's 80 percent failure rate.
This finding directly contradicts OpenAI's own claims about the new software. The company states that GPT-4 is 40 percent more likely to produce factual responses and 82 percent less likely to respond to disallowed content, highlighting a stark contrast between internal evaluations and independent testing.
The AI chatbot proves particularly dangerous because it formats its false narratives as highly convincing content, such as news articles, Twitter threads, and TV scripts. This advanced ability to mimic legitimate media formats makes the AI-generated misinformation much harder for average readers to identify and dismiss.